Using magnetic nanoparticles in MRI and hyperthermia treatment

Medical treatments using magnetic nanoparticles have already shown promising results in clinical practice. As an example, Feridex® and Resovist® have been approved as magnetic resonance imagining (MRI) contrast agents for liver imagining. A number of other magnetic nanoparticles for MRI have undergone different stages of clinical trials. The main reason for using magnetic nanoparticles as MRI contrast agents is that high-resolution images of different tissues enable accurate lesion detection. Moreover, magnetic nanoparticles have recently been used in clinical trials for hyperthermia treatment of glioblastoma, prostate and pancreatic cancer because under high temperature cells undergo heat stress so apoptosis, signal transduction and protein expression can be controlled by adjusting size, shape, surface, composition of materials and parameters of applied electromagnetic field.
Detailed information can be found in the relevant review articles on the application of magnetic nanoparticles in MRI and hyperthermia.


Descriptors

Dataset was formed during the data processing that includes implementation of new descriptors using initial parameters, correlation matrix analysis, removing outliers and filling of missing data.
SAR and r1, r2 relaxivities were chosen as predictive parameters because they reflect performance of nanoparticles in hyperthermia treatment and MRI. They depend on many factors, so essential parameters were extracted from articles and using them some derivative descriptors were calculated. All descriptors in the database reflect the composition of core as average magnetic moment of core metals; composition of surface as spin of surface atoms, number of H acceptors and LogP for organic coating molecules; shape as maximum to minimum length ratio; size and surface fraction as surface area volume ratio; magnetic properties of nanoparticle as coercivity, remanence and saturation magnetization; experimental conditions as field strength, amplitude and frequency, concentration of nanoparticles.




Prediction algorithms

For SAR prediction LightGBM Regressor was used with Q2 = 0.86, RMSE = 0.28 in 10-fold cross validation and R2 = 0.86, RMSE = 0.26 on test samples; ExtraTrees Regressor showed Q2 = 0.72, RMSE = 0.30 in 10-fold cross validation and R2 = 0.78, RMSE = 0.27 on test samples for prediction r1 relaxivity and this algorithm also is the best for prediction r2 relaxivity values – it showed Q2 = 0.71 and RMSE = 0.25 in 10-fold cross-validation and R2 = 0.75 and RMSE = 0.22 for test sample


Resource functions

— Database of magnetic nanoparticles for r1/r2 relaxivities and SAR with samples mined from scientific articles and their visualizations
— The opportunity to offer your sample
— Prediction algorithms for different levels of user request:
– The basic level (MRI and hyperthermia): user enters only the composition of core and size of nanoparticles. When predicting it is considered that the nanoparticle is spherical and uncoated
– The progressive level (MRI and hyperthermia) suggests describing nanoparticle: its core and shell composition, size and shape as well as experimental conditions, except of magnetic properties of nanoparticles such as saturation magnetization, remanence and coercivity
– Advanced level (MRI and hyperthermia) assumes full description of a system


Prediction limitations

Every machine learning model has its limitations that are important to keep in mind to avoid or minimize misuse. It should be noticed, that despite high accuracy of created ML algorithms, data that is used in our models is limited by nanoparticles studied in scientific articles as well as number of provided experiments. Thus, created models reproduce only those properties that are contained in our database. For instance, in our case, most of the data in the database refers to ferrites with different compositions, so prediction of properties of other types of materials may be inaccurate. Moreover, descriptors used in this work do not reflect all aspects of a given system. For instance, the lack of available data for hydrodynamic radius and zeta-potential prevented us from including this information in the training set, thus limiting the capability of the model to explain behavior of nanoparticles in the aqueous environment. Shell thickness is an important parameter of magnetic nanoparticles that is also missing in the training set due to very few articles reporting its quantification. It is noteworthy that all the data used to train the ML models was collected from in vitro studies, so our models do not reflect behavior of magnetic nanoparticles in vivo at all. For prediction in more complex biological systems, a much larger dataset covering more descriptors obtained from the relevant in vivo studies is required.